Tian Dong (Shanghai Jiao Tong University), Minhui Xue (CSIRO's Data61), Guoxing Chen (Shanghai Jiao Tong University), Rayne Holland (CSIRO's Data61), Yan Meng (Shanghai Jiao Tong University), Shaofeng Li (Southeast University), Zhen Liu (Shanghai Jiao Tong University), Haojin Zhu (Shanghai Jiao Tong University)

Open-source Large Language Models (LLMs) have recently gained popularity because of their comparable performance to proprietary LLMs. To efficiently fulfill domain-specialized tasks, open-source LLMs can be refined, without expensive accelerators, using low-rank adapters. However, it is still unknown whether low-rank adapters can be exploited to control LLMs. To address this gap, we demonstrate that an infected adapter can induce, on specific triggers, an LLM to output content defined by an adversary and to even maliciously use tools. To train a Trojan adapter, we propose two novel attacks, POLISHED and FUSION, that improve over prior approaches. POLISHED uses a superior LLM to align naïvely poisoned data based on our insight that it can better inject poisoning knowledge during training. In contrast, FUSION leverages a novel over-poisoning procedure to transform a benign adapter into a malicious one by magnifying the attention between trigger and target in model weights. In our experiments, we first conduct two case studies to demonstrate that a compromised LLM agent can use malware to control the system (e.g., a LLM-driven robot) or to launch a spear-phishing attack. Then, in terms of targeted misinformation, we show that our attacks provide higher attack effectiveness than the existing baseline and, for the purpose of attracting downloads, preserve or improve the adapter’s utility. Finally, we designed and evaluated three potential defenses. However, none proved entirely effective in safeguarding against our attacks, highlighting the need for more robust defenses supporting a secure LLM supply chain.

View More Papers

Power-Related Side-Channel Attacks using the Android Sensor Framework

Mathias Oberhuber (Graz University of Technology), Martin Unterguggenberger (Graz University of Technology), Lukas Maar (Graz University of Technology), Andreas Kogler (Graz University of Technology), Stefan Mangard (Graz University of Technology)

Read More

User Comprehension and Comfort with Eye-Tracking and Hand-Tracking Permissions...

Kaiming Cheng (University of Washington), Mattea Sim (Indiana University), Tadayoshi Kohno (University of Washington), Franziska Roesner (University of Washington)

Read More

PolicyPulse: Precision Semantic Role Extraction for Enhanced Privacy Policy...

Andrick Adhikari (University of Denver), Sanchari Das (University of Denver), Rinku Dewri (University of Denver)

Read More

GadgetMeter: Quantitatively and Accurately Gauging the Exploitability of Speculative...

Qi Ling (Purdue University), Yujun Liang (Tsinghua University), Yi Ren (Tsinghua University), Baris Kasikci (University of Washington and Google), Shuwen Deng (Tsinghua University)

Read More